CN114757183B - Cross-domain emotion classification method based on comparison alignment network - Google Patents
Cross-domain emotion classification method based on comparison alignment network Download PDFInfo
- Publication number
- CN114757183B CN114757183B CN202210373901.XA CN202210373901A CN114757183B CN 114757183 B CN114757183 B CN 114757183B CN 202210373901 A CN202210373901 A CN 202210373901A CN 114757183 B CN114757183 B CN 114757183B
- Authority
- CN
- China
- Prior art keywords
- domain
- cross
- target
- original
- field
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 230000008451 emotion Effects 0.000 title claims abstract description 65
- 238000000034 method Methods 0.000 title claims abstract description 35
- 230000006870 function Effects 0.000 claims description 32
- 238000012549 training Methods 0.000 claims description 24
- 238000013145 classification model Methods 0.000 claims description 14
- 238000012546 transfer Methods 0.000 claims description 10
- 238000012360 testing method Methods 0.000 claims description 8
- 238000012795 verification Methods 0.000 claims description 8
- 238000011156 evaluation Methods 0.000 claims description 6
- 238000007781 pre-processing Methods 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 3
- 230000011218 segmentation Effects 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 2
- 238000012545 processing Methods 0.000 claims description 2
- 238000010200 validation analysis Methods 0.000 claims description 2
- 238000004458 analytical method Methods 0.000 abstract description 7
- 238000003062 neural network model Methods 0.000 abstract description 5
- 238000011160 research Methods 0.000 abstract description 3
- 239000000284 extract Substances 0.000 abstract description 2
- 238000003058 natural language processing Methods 0.000 abstract description 2
- 230000000694 effects Effects 0.000 description 7
- 230000006978 adaptation Effects 0.000 description 5
- 238000012800 visualization Methods 0.000 description 5
- 238000010606 normalization Methods 0.000 description 4
- 238000002372 labelling Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 238000010411 cooking Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000011065 in-situ storage Methods 0.000 description 1
- 238000013508 migration Methods 0.000 description 1
- 230000005012 migration Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000008450 motivation Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/35—Clustering; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Machine Translation (AREA)
Abstract
The invention relates to a cross-domain emotion analysis method based on a comparison alignment network, and belongs to the technical field of fine-granularity emotion analysis in natural language processing. The invention researches an underexplored scene of cross-domain emotion classification, namely a scene with a few samples in the target domain. In this scenario, the present invention proposes a neural network model named Contrast Alignment Network (CAN). The model first randomly extracts two instances from the original domain and the target domain, and then trains them according to the instances of the combined target domain and the original domain. The first objective is to minimize classification errors on the original domain. The second is a pairwise comparison objective, where the distance measure between the objective domain instance and the original domain instance in a pair is minimized if they express the same emotion, otherwise the measure is maximized at a constant upper limit. The method solves the problem of limited data resources in the target field in the cross-field emotion classification task, and improves the use experience of users.
Description
Technical Field
The invention relates to a cross-domain emotion classification method, in particular to a cross-domain emotion analysis method based on a comparison alignment network, and belongs to the technical field of fine-granularity emotion analysis in natural language processing.
Background
Cross-domain emotion classification (Cross Domain Sentiment Classification, CDSC) is an important task aimed at transferring learned knowledge from the original domain to the target domain. CDSC enables emotion classification models trained in the original domain with a large amount of labeled data to perform well in target domain data with limited training samples. This situation is common in the industry when the data of the target domain is lacking and the data of the original domain is sufficient, and has a challenge, the main challenge being domain transfer (or distribution transfer) between the source domain and the target domain. The domain transfer problem is mainly a distribution difference between any two domains, for example, a word used in the medical domain is greatly different from a word used in the restaurant domain.
Domain transfer is an important problem in cross-domain emotion classification, which can be alleviated to a large extent by domain adaptation methods. Currently, researchers have proposed various field-adaptive models. These models require a large amount of unlabeled data from the target domain so that they can learn a good representation of each target instance as the correct input to the classifier trained in the original domain.
Currently, unsupervised Domain Adaptation (UDA) techniques have been used to solve the domain transfer problem. Essentially, the UDA utilizes other unlabeled data in the target domain to minimize domain offset by aligning statistics across domains. However, in practical applications, a large amount of unlabeled target area data required for UDA may not be sufficiently available, thereby limiting the applicability of UDA.
Meanwhile, there is a scene which is not fully explored at present in cross-domain emotion classification, namely, a scene with a few samples in the target domain, and the scene exists in many practical applications in the industry. Unlike an unsupervised scenario, a few sample scenario does not require additional unlabeled target domain data, but relies only on the scarce labeling data available in the target domain.
Disclosure of Invention
The invention aims to overcome the defects of the prior art, creatively provides a cross-domain emotion classification method based on a comparison alignment network, and aims to solve the technical problem of data resource limitation in the target domain. The invention researches an underexplored scene of cross-domain emotion classification, namely a scene with a few samples in the target domain. Unlike an unsupervised scenario, a few sample scenario does not require additional unlabeled target domain data, but relies only on the scarce labeling data available in the target domain. In this scenario, the present invention proposes a neural network model named Contrast Alignment Network (CAN). The model first randomly extracts two instances from the original domain and the target domain, and then trains them according to the instances of the combined target domain and the original domain. The first objective is to minimize classification errors on the original domain. The second is a pairwise comparison objective, where the distance measure between the objective domain instance and the original domain instance in a pair is minimized if they express the same emotion, otherwise the measure is maximized at a constant upper limit.
Notably, the motivation for the alignment of networks is not just to address resource limitation issues. In fact, by deducing that the model provided by the invention has more complete learning limit, theoretical basis is provided for the basic efficacy of the model.
Cross-domain emotion analysis CDSC requires an in-situ domainModel of middle training, original fieldContains n s examples, which are specific to the target field/>And performing emotion classification. X i s is the ith sample sentence in the original field, and y i s is the label corresponding to the ith sample in the original field. In other words, the original field/>In example sample distribution compliance/>Also, there is/>, on the target areaX t represents a sample sentence of the target field t, and y t represents a tag corresponding to the sample sentence of the target field t. /(I)And/>Representing the domain offset between the joint probabilities of the source and target, respectively, and
The concept of domain adaptation is imposed to alleviate the domain transfer problem. In the case that there is insufficient tagged data in the target domain, the unsupervised domain adaptation method uses a large amount of additional untagged target dataTo align the two fields, n t being the target field and the number of marked data samples. However, the large amount of unlabeled data required may not always be available in the target area. Thus, the method focuses on a more realistic case of cross-domain emotion analysis, i.e., a small sample case, instead using a small number of auxiliary markers to target domain data/>N t＜＜ns,nt is the number of auxiliary mark target field data samples.
Inspired by the adaptation work of the existing unsupervised field, the learning target is decomposed into two parts: and calculating the discriminative original domain risk and regularized domain transfer. In order to further minimize the risk of the original domain, the invention provides a method for comparing and aligning the original domain and the target domain by taking example classification information as a condition.
The technical mode adopted by the invention is as follows.
An emotion classification method based on a comparison alignment network comprises the following steps:
First, text data preprocessing is performed.
And loading a comment corpus and a pre-training language model, and carrying out text preprocessing and text data formatting on comment text data in the comment corpus. Wherein, the pre-training language model can adopt BERT model, roBERTa model and the like.
Then, constructing a cross-domain emotion classification model based on the comparison alignment network.
The cross-domain emotion classification model f based on the contrast alignment network provided by the invention comprises an encoder g θ and a classifier h φ, and on the basis of the architecture, an original domain classification target loss function is introducedAnd comparing the target domain classification target loss function/>
Wherein the encoder g θ uses the pre-trained language model as a basis for encoding context information of comment sentences. Preferably, the encoder uses CLS (sentence vector representation) full sentence representation of the pre-trained language model as the context hidden state representation vector H of the whole comment sentence, h= { H 1,h2,...,hn},hn representing the hidden state representation vector of the nth token.
The classifier h φ consists of a multi-layer perceptron MLP and a softmax layer (soft maximization normalization layer). Wherein, multilayer perceptron includes four layers, does in proper order: a full connection layer, a ReLU (Linear rectifier function) activation function layer, a dropout layer (random discard layer) and a full connection layer. The output representation through the MLP is sent to the softmax layer (soft maximization normalization layer), from which the corresponding losses are calculated.
Then, the discriminative primitive domain risk is calculated.
In the method, for the discriminative original domain risk, an empirical classification loss term of the original domain is adopted, and a classification target is modeled as a loss based on cross entropy
Wherein n s is the number of data samples in the original field; y i is the label of the ith sample of the original field data,And (5) predicting labels on the ith sample of the original field data for the model.
Then, the original domain and the target domain are aligned by comparing the example-level classification information.
The goal is to minimize domain offset under limited target domain data. Although alignment of distribution levels is difficult to achieve with limited target data, alignment of instance levels may be combined and alignment of instance levels may achieve distribution level alignment with some probability.
In the method of the present invention, contrast loss is introduced to minimize the distance metric between the sample pair (if they share the same emotion) between the original domain instance and the target domain instance, otherwise the metric is maximized. Therefore, regardless of the field, emotion tags can be accurately assigned. Although the distance measure between the original domain and the target domain can be directly applied to the input x, the present invention follows most domain-adaptive methods to align the hidden representation Z, which will be more abstract and can capture more semantic information.
Specifically, given any pair ofSpecific contrast loss/>The calculation is as follows:
Wherein, X i s represents the ith sample comment sentence in the original field, X j t represents the jth sample comment sentence in the target field, y i s represents the label corresponding to the ith sample comment sentence in the original field, and y j t represents the label corresponding to the jth sample comment sentence in the target field; representing a distance measure between the original domain instance and the target domain instance; /(I) Indicating a function for the equation; m is a predefined constant.
The contrast loss objective pushes the instances of the same emotion polarity closer, rejecting the instances of different emotion polarity. Furthermore, the goal is not to push different clusters to infinity, but rather to limit the scope of rejection to a constant as a slack in the learning algorithm.
And then, carrying out regularized field transfer.
The overall objective of the method includes cross entropy loss function of the original domain dataComparing and aligning original domain and target domain loss function/>And pass/>Regularization minimizes cross entropy loss function/>, of raw domain dataAnd comparing the aligned original domain with the target domain loss function/>
Wherein the overall objective function is:
Where α is a trade-off term between classification and comparison targets and λ is the regularization coefficient of all model parameters Θ= { θ, Φ }.
Then, model training is performed. For the overall objective function using a standard batch random gradient descent algorithmTraining is performed.
Specifically, batch iterative training is carried out on all training samples in a training set, and a trained cross-domain emotion classification model based on a comparison alignment network is obtained.
And finally, performing cross-domain emotion classification by using a trained cross-domain emotion classification model based on the comparison and alignment network.
Advantageous effects
Compared with the prior art, the method has the following advantages:
1. The method solves the problem that data resources in the target field are limited in the cross-field emotion classification task. The invention researches an underexplored scene of cross-domain emotion classification, namely a scene with a few samples in the target domain. Unlike an unsupervised scenario, a few sample scenario does not require additional unlabeled target domain data, but relies only on the scarce labeling data available in the target domain.
2. The neural network model for the alignment network proposed by the method comprises two targets, wherein the first target is to minimize classification errors in the original field. The second is a pairwise comparison objective, where the distance measure between the objective domain instance and the original domain instance in a pair is minimized if they express the same emotion, otherwise the measure is maximized at a constant upper limit.
3. The contrast alignment network model provided by the invention has the advantage that the performance of the contrast alignment network model on the cross-domain emotion classification task is obviously superior to that of a corresponding baseline model. The performance of other baseline models on cross-domain datasets drops dramatically. In contrast, the proposed contrast aligned neural network model is more robust than other baseline models.
4. The contrast alignment neural network model provided by the invention not only solves the problem of limited resources in the target field, but also has a more complete learning limit, thereby providing a theoretical basis for the basic efficacy of the model.
5. The method improves the problems in cross-domain emotion classification of the existing fine-grained emotion analysis, and can well improve the use experience of users.
Drawings
FIG. 1 is an overall flow chart of the method of the present invention.
FIG. 2 is an example diagram of a cross-domain emotion classification task.
Fig. 3 is a visual representation of the effect of the method of the present invention.
Detailed Description
The invention is described in further detail below with reference to the drawings and examples.
Examples
As shown in fig. 1, a cross-domain emotion classification method based on a comparison alignment network includes the following steps:
Step 1: text preprocessing.
First, a comment corpus and a pre-trained language model are loaded. The pre-training language model may be a BERT model, or may be another model (such as RoBERTa models).
Then, text preprocessing and text data formatting are performed on the comment corpus.
Specifically, the method comprises the following steps:
Step 1.1: and extracting attribute words, viewpoint words and position information of the attribute words and the viewpoint words from each comment sentence.
Step 1.2: and pre-segmenting the comment sentences by using a nltk word segmentation device, and separating the segmented token words by using spaces.
Step 1.3: after the comment sentence is divided into a token sequence, two special token words are added: [ CLS ], [ SEP ], thereby constructed in the general input form: s= { [ CLS ], w 1,w2...,wn, [ SEP ] }, n represents the total number of token words of the comment sentence, and w n represents the n-th token of the comment sentence.
Step 1.4: text data formatting is performed.
And (3) carrying out filling processing on each comment sentence token word sequence to ensure that the length of each comment sentence token word sequence is 128. Each token word in the comment sentence is subjected to tokenize operations using the word segmenter tokenizer of the pre-trained language model. The processed data set is divided into a training set, a validation set and a test set, and a batch data form is constructed.
Step 2: and constructing a cross-domain emotion classification model based on the comparison alignment network.
The cross-domain emotion classification model f based on the comparison alignment network consists of an encoder g θ and a classifier h φ. Based on encoder g θ and classifier h φ architecture, the original domain classification target loss function is introducedAnd comparing the target domain classification target loss function/>
Wherein the encoder g θ encodes context information of the comment sentence. For emotion classification tasks, it is important that emotion information of a sentence is contained in the context of the sentence and that the entire sentence is context modeled. Thus, using the full sentence CLS representation of the pre-trained language model, the context hidden state representation vector H, h= { H 1,h2,...,hn},hn, represents the hidden state representation vector of the nth token as the entire comment sentence. Each token word in the sequence of comment sentences is mapped to a code vector.
The classifier h φ consists of a multi-layer perceptron MLP and a softmax layer (soft maximization normalization layer). Wherein, the multilayer perceptron includes four layers: full connection layer, reLU (linear rectification function) activation function layer, dropout layer (random discard layer) and full connection layer. The output representation through the fully-connected layer is sent to the softmax layer (soft maximization normalization layer), whereby the corresponding label is predicted and the corresponding target loss is calculated.
Step 3: and calculating the risk of the discriminant primitive field.
For the discriminative primitive domain risk, the invention adopts the empirical classification loss term of the primitive domain, and then models the classification target as the loss based on cross entropy
Wherein n s is the number of samples of the original field data, y i is the label of the ith sample of the original field data,And (5) predicting labels on the ith sample of the original field data for the model.
Step 4: and comparing and aligning the original domain with the target domain by using example-level classification information.
The domain offset is minimized under limited target domain data. Although alignment of distribution levels is difficult to achieve with limited target data, alignment of instance levels may be combined and alignment of instance levels may achieve distribution level alignment with some probability.
Thus, the present invention introduces contrast loss to minimize the distance metric between the sample pair original domain instance and the target domain instance if the original domain instance and the target domain instance share the same emotion, and otherwise maximize the metric. Regardless of the field, emotion tags can be accurately assigned.
Although the distance measure between the original domain and the target domain can be directly applied to the input x, the present invention follows most domain-adaptive methods, aligning the hidden representation Z. This is more abstract and more semantic information can be captured.
Specifically, given any pair ofThe specific contrast loss is calculated as:
Wherein, Representing a distance measure between the original domain instance and the target domain instance. /(I)The function is indicated for the equation. m is a predefined constant.
The contrast loss objective pushes the instances of the same emotion polarity closer, rejecting the instances of different emotion polarity. Furthermore, the goal is not to push different clusters to infinity, but rather to limit the scope of rejection to a constant as a slack in the learning algorithm.
Step 5: regularized domain transfer.
The overall objective of the method includes cross entropy loss function of the original domain dataComparing and aligning original domain and target domain loss function/>By/>Regularization minimizes cross entropy loss function/>, of raw domain dataAnd comparing the aligned original domain with the target domain loss function/>
The overall objective function is:
Where α is a trade-off term between classification and comparison targets, λ is a regularization coefficient of all model parameters Θ= { θ, Φ } θ represents a parameter of the encoder, Φ represents a parameter of the classifier.
Step 6: training a cross-domain emotion classification model based on a comparison alignment network.
For the overall objective function using a standard batch random gradient descent algorithmAnd (5) performing optimization training.
Specifically, batch iterative training is carried out on all training samples in a training set, and a trained cross-domain emotion classification model based on a comparison alignment network is obtained.
Step 7: and performing cross-domain emotion classification by using the trained cross-domain emotion classification model based on the comparison and alignment network.
Further, the method can be evaluated. After the training is completed in the training set, a verification test is performed in the verification set used. The evaluation indexes used include:
for cross-domain emotion classification, using accuracy and F1 values as evaluation indexes;
And updating the optimal model for each round of verification and saving.
Test verification
The method was tested. Firstly, loading the parameters and test data of the optimal model stored before, and then converting the test data into a required format and inputting the required format into the optimal model for testing. Wherein, the evaluation index is the same as the evaluation index used in verification.
As shown in fig. 2, for the original field being the comment field of the notebook computer, the comment sentence "the brand notebook computer is too stuck, the use experience is extremely bad", the target field being the comment sentence of the restaurant field "the cooking of his home is general, the service is also general", and there is a fine-grained field deviation between the original field and the target field data, which makes most of the inter-field emotion classification methods and the unsupervised self-adaptive methods unable to perform another migration well, and cannot effectively judge the emotion polarity of the comment sentence of the target field on the target field.
As shown in fig. 3, the cross-domain (notebook computer comment field and restaurant comment field) effect visualization representation of the emotion classification method based on the contrast alignment network, the left graph is the cross-domain effect visualization of the non-contrast alignment network, and the right graph is the cross-domain effect visualization of the contrast alignment network. It can be seen that the model effect visualization with right graph with contrast aligned network has a more elegant manifold than the effect visualization without contrast aligned network, which means that the contrast aligned network has generalization capability even if the target data is small.
The foregoing is a preferred embodiment of the present invention and the present invention should not be limited to the embodiment and the disclosure of the drawings. All equivalents and modifications that come within the spirit of the disclosure are desired to be protected.
Claims (5)
1. The cross-domain emotion classification method based on the comparison alignment network is characterized by comprising the following steps of:
step 1: loading a comment corpus and a pre-training language model, and carrying out text preprocessing and text data formatting on comment text data in the comment corpus;
step 2: constructing a cross-domain emotion classification model based on a comparison alignment network;
The cross-domain emotion classification model f based on the comparison alignment network consists of an encoder g θ and a classifier h φ, and an original domain classification target loss function is introduced on the basis of the framework And comparing the target domain classification target loss function/>
The encoder g θ uses the pre-trained language model as a basis for encoding context information of comment sentences; the classifier h φ consists of a multi-layer perceptron MLP and a softmax layer; passing the output representation through the multi-layer perceptron to a softmax layer, thereby calculating a corresponding loss;
step 3: calculating the risk of the discriminant primitive field;
for discriminative primitive domain risk, modeling a classification target as a cross entropy-based loss by using an empirical classification loss term of the primitive domain
Wherein n s is the number of data samples in the original field; y i is the label of the ith sample of the original field data,A predictive label of the model on an ith sample of the original field data;
step 4: comparing and aligning the original field with the target field according to the example-level classification information;
Given any pair of Specific contrast loss/>The calculation is as follows:
wherein, X i s represents the ith sample comment sentence in the original field, X j t represents the jth sample comment sentence in the target field, y i s represents the label corresponding to the ith sample comment sentence in the original field, and y i t represents the label corresponding to the jth sample comment sentence in the target field; representing a distance measure between the original domain instance and the target domain instance; /(I) Indicating a function for the equation; m is a predefined constant;
step 5: performing regularization field transfer; the overall objective includes a cross entropy loss function of the raw domain data Comparing and aligning original domain and target domain loss function/>And pass/>Regularization minimizes cross entropy loss function of raw domain dataAnd comparing the aligned original domain with the target domain loss function/>
Wherein the overall objective function is:
where α is a trade-off term between classification and comparison targets, λ is the regularization coefficient of all model parameters Θ= { θ, Φ };
step 6: for the overall objective function using a standard batch random gradient descent algorithm Training to obtain a trained cross-domain emotion classification model based on a comparison alignment network;
step 7: and performing cross-domain emotion classification by using the trained cross-domain emotion classification model based on the comparison and alignment network.
2. The cross-domain emotion classification method based on a contrast alignment network of claim 1, wherein the encoder uses CLS full sentence representation of the pre-trained language model as context hidden state representation vector H of the whole comment sentence, h= { H 1,h2,…,hn},hn represents hidden state representation vector of the nth token.
3. The cross-domain emotion classification method based on a comparison alignment network of claim 1, wherein in step 2, the multi-layer perceptron comprises four layers, in order: a full connection layer, a ReLU activation function layer, a dropout layer and a full connection layer.
4. The cross-domain emotion classification method based on a comparison alignment network of claim 1, wherein step 1 comprises the steps of:
step 1.1: extracting attribute words, viewpoint words and position information of the attribute words and the viewpoint words from each comment sentence;
step 1.2: pre-segmenting the comment sentences by using nltk word segmentation devices, and separating segmented token words by using spaces;
Step 1.3: after the comment sentence is divided into a token sequence, two special token words are added: [ CLS ], [ SEP ], thereby constructed in the general input form: s= { [ CLS ], w 1,w2…,wn, [ SEP ] }, n represents the total number of token words of the comment sentence, and w n represents the n-th token of the comment sentence;
Step 1.4: formatting text data;
Performing filling processing on each comment sentence token word sequence to ensure that the length of each comment sentence token word sequence is 128; using a word segmentation device tokenizer of the pre-training language model to carry out tokenize operation on each token word in the comment sentence; the processed data set is divided into a training set, a validation set and a test set, and a batch data form is constructed.
5. The cross-domain emotion classification method based on a comparison alignment network of claim 1, wherein after training is completed in the training set, a verification test is performed in the verification set used, and the evaluation index includes:
for cross-domain emotion classification, using accuracy and F1 values as evaluation indexes;
Updating and saving the optimal model for each round of verification.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210373901.XA CN114757183B (en) | 2022-04-11 | 2022-04-11 | Cross-domain emotion classification method based on comparison alignment network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210373901.XA CN114757183B (en) | 2022-04-11 | 2022-04-11 | Cross-domain emotion classification method based on comparison alignment network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114757183A CN114757183A (en) | 2022-07-15 |
| CN114757183B true CN114757183B (en) | 2024-05-10 |
Family
ID=82328537
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210373901.XA Active CN114757183B (en) | 2022-04-11 | 2022-04-11 | Cross-domain emotion classification method based on comparison alignment network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114757183B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115239937B (en) * | 2022-09-23 | 2022-12-20 | 西南交通大学 | Cross-modal emotion prediction method |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109308318A (en) * | 2018-08-14 | 2019-02-05 | 深圳大学 | Training method, device, equipment and the medium of cross-domain texts sentiment classification model |
| CN109753566A (en) * | 2019-01-09 | 2019-05-14 | 大连民族大学 | The model training method of cross-cutting sentiment analysis based on convolutional neural networks |
| CN110032646A (en) * | 2019-05-08 | 2019-07-19 | 山西财经大学 | The cross-domain texts sensibility classification method of combination learning is adapted to based on multi-source field |
| CN110503092A (en) * | 2019-07-22 | 2019-11-26 | 天津科技大学 | The improvement SSD monitor video object detection method adapted to based on field |
| AU2020100710A4 (en) * | 2020-05-05 | 2020-06-11 | Chen, Dadu Mr | A method for sentiment analysis of film reviews based on deep learning and natural language processing |
| CN112163091A (en) * | 2020-09-25 | 2021-01-01 | 大连民族大学 | CNN-based aspect-level cross-domain emotion analysis method |
| CN113326378A (en) * | 2021-06-16 | 2021-08-31 | 山西财经大学 | Cross-domain text emotion classification method based on parameter migration and attention sharing mechanism |
-
2022
- 2022-04-11 CN CN202210373901.XA patent/CN114757183B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109308318A (en) * | 2018-08-14 | 2019-02-05 | 深圳大学 | Training method, device, equipment and the medium of cross-domain texts sentiment classification model |
| CN109753566A (en) * | 2019-01-09 | 2019-05-14 | 大连民族大学 | The model training method of cross-cutting sentiment analysis based on convolutional neural networks |
| CN110032646A (en) * | 2019-05-08 | 2019-07-19 | 山西财经大学 | The cross-domain texts sensibility classification method of combination learning is adapted to based on multi-source field |
| CN110503092A (en) * | 2019-07-22 | 2019-11-26 | 天津科技大学 | The improvement SSD monitor video object detection method adapted to based on field |
| AU2020100710A4 (en) * | 2020-05-05 | 2020-06-11 | Chen, Dadu Mr | A method for sentiment analysis of film reviews based on deep learning and natural language processing |
| CN112163091A (en) * | 2020-09-25 | 2021-01-01 | 大连民族大学 | CNN-based aspect-level cross-domain emotion analysis method |
| CN113326378A (en) * | 2021-06-16 | 2021-08-31 | 山西财经大学 | Cross-domain text emotion classification method based on parameter migration and attention sharing mechanism |
Non-Patent Citations (1)
| Title |
|---|
| 基于集成深度迁移学习的多源跨领域情感分类;赵传君;王素格;李德玉;;山西大学学报(自然科学版);20180404(第04期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114757183A (en) | 2022-07-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112214995B (en) | Hierarchical multitasking term embedded learning for synonym prediction | |
| WO2023065545A1 (en) | Risk prediction method and apparatus, and device and storage medium | |
| CN111738003B (en) | Named entity recognition model training method, named entity recognition method and medium | |
| CN111737552A (en) | Method, device and equipment for extracting training information model and acquiring knowledge graph | |
| CN112541355B (en) | Entity boundary type decoupling few-sample named entity recognition method and system | |
| CN112084327A (en) | Classification of sparsely labeled text documents while preserving semantics | |
| CN111985239A (en) | Entity identification method and device, electronic equipment and storage medium | |
| CN111506732B (en) | Text multi-level label classification method | |
| CN111597340A (en) | Text classification method and device and readable storage medium | |
| CN113254675B (en) | Knowledge graph construction method based on self-adaptive few-sample relation extraction | |
| CN112100348A (en) | Knowledge base question-answer relation detection method and system of multi-granularity attention mechanism | |
| CN114912423A (en) | Method and device for analyzing aspect level emotion based on transfer learning | |
| CN114841151B (en) | Medical text entity relation joint extraction method based on decomposition-recombination strategy | |
| CN116383399A (en) | Event public opinion risk prediction method and system | |
| CN115063119A (en) | Recruitment decision system and method based on adaptivity of recruitment behavior data | |
| CN113704392A (en) | Method, device and equipment for extracting entity relationship in text and storage medium | |
| CN115935991A (en) | Multitask model generation method and device, computer equipment and storage medium | |
| CN113591955A (en) | Method, system, equipment and medium for extracting global information of graph data | |
| CN115563314A (en) | Knowledge graph representation learning method for multi-source information fusion enhancement | |
| CN116861924A (en) | Project risk early warning method and system based on artificial intelligence | |
| CN114757183B (en) | Cross-domain emotion classification method based on comparison alignment network | |
| CN115659947A (en) | Multi-item selection answering method and system based on machine reading understanding and text summarization | |
| CN115017879A (en) | Text comparison method, computer device and computer storage medium | |
| CN114780723A (en) | Portrait generation method, system and medium based on guide network text classification | |
| CN114582448A (en) | Epidemic case information extraction framework construction method based on pre-training language model |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |